A Long-Term Gesture Recognition Method Based on Surface Electromyography Signals

Abstract

The present invention relates to a long-term gesture recognition method based on surface electromyography signals. It includes two parts: the first part is data preprocessing, including feature extraction and feature dimensionality reduction; the second part is the adaptive update of the classification model. First is the data preprocessing part. Differential common spatial pattern (DCSP) features are extracted from the gesture data of existing days (usually the data collected on the first day), and then the features are transformed by the nonnegative matrix factorization (NMF) algorithm with activation coefficient normalization. Then is the adaptive update part of the classification model. The first repeated experiment of each test day is used as unlabeled samples, and these samples are evaluated and labeled by the Clustering and classification self-training (CCST) method. The qualified samples, labels and existing data are used to retrain the classification model together. The present invention avoids the trouble of users collecting data for calibration every day and improves the user's comfort level.

Description

Technical Field

[0001] The present invention relates to a long-term gesture recognition method based on surface electromyogram signals. Background Art

[0002] Surface electromyogram (sEMG) signals are superimposed electrical signals formed on the skin surface by the motor unit action potentials (MUAPs) of movement-related muscles propagating along the muscle fiber direction. Compared with invasive electromyogram signals, the acquisition method of surface electromyogram signals is simple and the process is non-invasive. On the one hand, surface electromyogram signals are considered to contain a large amount of user movement information, and information such as muscle contraction force and joint torque can be decoded from them, which is widely used to predict limb states or predict continuous joint information. On the other hand, surface electromyogram signals have the "electromechanical delay" (EMD) characteristic. There is a certain delay between the generation of surface electromyogram signals and the execution of actual actions, which helps to predict the user's movement intention. The above advantages make the human-computer interaction system developed based on surface electromyogram signals more and more common in daily life and develop in the direction of being portable and wearable.

[0003] Currently, the human-computer interaction system based on surface electromyogram signals usually includes two processes: offline processing and online analysis. In the offline processing stage, first, electromyogram signals with known action labels of users are collected in a laboratory environment, and then through preprocessing and feature extraction. Finally, the action features and their corresponding labels jointly complete the training of the classifier. During online analysis, when an electromyogram signal of an unknown action is obtained, it also needs to go through the steps of preprocessing and feature extraction, and then be sent to the classifier obtained in the offline processing stage to output the category corresponding to this segment of electromyogram signal in real time. After decades of research, the current action recognition has developed from the initial binary classification to predicting dozens of hand and wrist actions, and its classification accuracy basically exceeds 90%.

[0004] However, a common feature of these studies is that their collection time in a limited laboratory environment is short (i.e., once or several times a day), and the performance of the model has not been verified on gesture data over a long period of time. In fact, regardless of the performance of the model at the time of creation, the non-stationarity of electromyogram signals will gradually reduce the model performance over time (up to 20-30%). In order to ensure the recognition accuracy, the electromyogram control system often needs to regularly re-collect data to re-train the classifier during long-term use, which brings great trouble to consumers and has become the main obstacle to the commercialization of consumer-grade electromyogram control systems.

[0005] In recent years, scholars have proposed many solutions to the effects of non-stationarity of EMG signals. These studies can be divided into three main research categories. The first category is to explore the nature of temporal changes to seek answers to the following questions: (1) when changes occur, (2) the impact of these changes on classifier performance (when no cyclic training is performed), and (3) the amount and stratification of training data required to achieve stable parameter estimation. The second category is the identification of stable classification, feature extraction, and dimensionality reduction algorithms that are invariant to changes over time. The third category is to develop adaptive algorithms that can actively retrain and / or relabel classification models.

[0006] In the above literature, firstly, in the long-term gesture recognition process, only the inherent time-varying characteristics of the electromyographic signal are taken into account. In fact, since the gesture recognition system of the electromyographic signal will eventually be applied to wearable devices, the device needs to be put on and taken off many times during long-term use, which makes the electrode offset also become an important reason affecting the accuracy of long-term gesture recognition. In the gesture collection scheme of this method, in the gesture recognition of 30 consecutive days, the collection armband of each day is located in one of the offset positions (clockwise deviation from the neutral position 8mm, neutral position, counterclockwise deviation from the neutral position 8mm), which is closer to the actual application. Secondly, the three types of methods have their own defects. The first type of method informs the user of the number of pre-training days required for long-term gesture recognition by studying the time characteristics of electromyographic gesture recognition, but it increases the burden of user training. The second type of method does not need to collect data again during the gesture recognition process, but the first and second types of methods have a problem that the methods are based on the data that already contains the number of days, and their robustness to the unknown new day is relatively poor, and the recognition accuracy is generally difficult to meet the requirements. The third type of method is more stable, but in order to ensure recognition accuracy, most studies still require a calibration set with known labels to achieve fast calibration of the classifier, which brings the burden of daily retraining for users. In summary, the design of long-term gesture recognition solutions should focus on reducing the burden of user training and retraining and achieving the classification accuracy required by practical applications. Summary of the invention

[0007] The purpose of the present invention is to provide a long-term gesture recognition method based on surface electromyography signals, which solves the influence of the inherent time-varying characteristics of electromyography signals and electrode offset on the accuracy of gesture recognition, adopts an unsupervised method to update the classification model, avoids the trouble of users collecting data for calibration every day, and improves the user's comfort.

[0008] To achieve the above object, the technical solution of the present invention is: a long-term gesture recognition method based on surface electromyography signal, comprising:

[0009] Data preprocessing: Extract differential common spatial pattern (DCSP) features from the gesture data of existing days, and transform the DCSP features through the non-negative matrix factorization (NMF) algorithm with activation coefficient normalization;

[0010] Adaptive update of the classification model: Use the first repeated experiment of each test day as unlabeled samples, evaluate and label the unlabeled samples using the classification and clustering self-training (CCST) method, and retrain the classification model with the qualified samples, labels, and existing data together.

[0011] In an embodiment of the present invention, the specific implementation of extracting the differential common spatial pattern (DCSP) features is as follows:

[0012] Calculate the average covariance matrix of each class signal through formula (1):

[0013]

[0014] where i = 1, 2, 3... c, n is the number of sampling points of the i-th type of action, and X i represents the original electromyogram signal of the i-th category.

[0015] By solving formula (2), find an optimal projection matrix w such that the difference between the i-th class and the j-th class is the largest in the corresponding direction:

[0016] w×R i =λ×w×R j (2)

[0017] Calculate an optimal projection matrix for each pair of all classes, combine all projection matrices, and project the original electromyogram signal X, as shown in formula (3):

[0018] A=WX (3)

[0019] Then, solve the first-order and second-order difference signals of the projected signal and take the variance as shown in (4) to obtain A 0 、A 1 、A 2 :

[0020] A 0 =var(A) A 1 =var(ΔA) A 2 =var(Δ 2 A) (4)

[0021] var() represents calculating the variance of the signal, and Δ n represents calculating the n-th derivative of the signal, and then perform a power exponential transformation through formula (5) to obtain B 0 、B 1 、B 2, λ is usually taken as 0.1, where equation (6) combines the features, and finally logarithmic normalization is performed through (7) to obtain the DCSP features:

[0022]

[0023] C = [B 0 B 1 B 2 (6)

[0024] D = log(C / sum(C)) (7).

[0025] In an embodiment of the present invention, the specific implementation manner of converting the DCSP features through the non - negative matrix factorization algorithm NMF with activation coefficient normalization is as follows:

[0026] The method of combining the NMF algorithm with activation coefficient normalization is named the NMFH algorithm. The DCSP features are converted through the NMFH algorithm to increase the distinguishability of different actions of the DCSP features in long - time gesture recognition;

[0027] After the DCSP features are extracted, the features of the training day are named V, and the features of the test day are named V'. V is decomposed into the product of the collaborative matrix W ∈ R M×K and the activation coefficient matrix H ∈ R K×N Specifically, the element in the m - th row and n - th column of V can be decomposed as:

[0028]

[0029] where W is the muscle collaboration matrix, H is the activation matrix, M is the number of muscles, N is the number of feature samples, and K is the dimension of the feature after dimensionality reduction. First, W and H are initialized as random non - negative matrices, taking K = 10. Then, each element in W and H is iteratively solved through equations (9) - (11), where equation (11) is to add the normalization of the activation coefficient H during the iteration:

[0030]

[0031]

[0032]

[0033] The test set V' fixes the W matrix obtained by iterating the training set, initializes H' as a random non - negative matrix, and obtains the test feature H' through equations (10) - (11).

[0034] In an embodiment of the present invention, the specific implementation steps of the adaptive update of the classification model are as follows:

[0035] Step 1: Initialization of the classification model and acquisition of classification labels:

[0036] Train an initial classification model with the samples of the training day. The first repeated experiment on each test day is marked by the initial classification model as unlabeled samples to obtain classification labels and classification probabilities, and the remaining repeated experiments are used as the test set to verify the performance of the algorithm;

[0037] Step 2: Acquisition of clustering labels based on fuzzy C-means clustering:

[0038] Obtain the clustering labels of unlabeled samples through the fuzzy C-means clustering algorithm. Divide a set containing n data objects into c classes, and its clustering centers are represented by c i The objective function is:

[0039]

[0040] In the formula, u ij takes values between 0 and 1, c i represents the clustering center of the i-th class, d ij represents the Euclidean distance between the i-th clustering center c i and the j-th data object x j m ∈ [1, +∞) is the weighted exponent used to control the fuzziness of clustering; the larger the m value, the greater the fuzziness. Under the constraints of the constraint conditions, use the Lagrange multiplier method to obtain the iterative expressions of the clustering center and membership degree for minimizing formula (12) as shown in (13) and (14):

[0041]

[0042]

[0043] Take the classification probability of the unlabeled samples in Step 1 as the initial membership degree matrix, and continuously iterate through formulas (13) and (14) to output the clustering center and membership degree matrix of each round until the objective function (12) reaches the minimum value or exceeds the set maximum number of iterations, and then the iteration ends; output the membership degree matrix, and take the class corresponding to the maximum value of the membership degree in each unlabeled sample as the clustering label of this unlabeled sample;

[0044] Step 3: Sample evaluation and classifier update:

[0045] Compare the classification labels and clustering labels of each sample, and use the samples with consistent classification labels and clustering labels to update the LDA classifier. The LDA classifier only needs to update two parameters to achieve the update of the classifier; the update steps are as follows:

[0046] μ 2= rμ 0 + (1 - r)μ 1

[0047] Σ 2 = rΣ 0 + (1 - r)Σ 1

[0048] where u 2 , u 0 and u 1 represent the average values of updated data, training data, and calibration data respectively; Σ 2 , Σ 0 and Σ 1 represent the covariance matrices of updated data, training data, and calibration data respectively; r is a balance coefficient with a value range between 0 and 1;

[0049] Step 4, Cascade Adaptation:

[0050] Adopt a cascade adaptation method, that is, use the training set and the selected data to generate a classification model on the first day of testing, and use the updated model on the first day and the newly selected data weighted to generate the classification model on the second day of testing. And every day after that, update based on the model updated the previous day.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] First, adopt the DCSP feature combined with the NMFH algorithm for feature transformation, which improves the discrimination between different actions, reduces the dimension of the original features, and reduces the computational amount.

[0053] Second, propose the CCST adaptive learning method, screen and label the unlabeled samples that have been output, retrain the classifier to make the classification model adaptable to changes, and adopt an unsupervised update method, avoiding the trouble of still needing to collect data for calibration every day.

[0054] Third, the features and the adaptive algorithm are verified on a longer-term electromyogram gesture data set (electromyogram data continuously collected for 30 days), and the daily data is also at different electrode offset positions, which is more in line with the situation where the electromyogram device needs to be taken off and worn during long-term use.

[0055] The product of the method of the present invention can be applied to all human-computer interaction systems based on surface electromyogram signals that require long-term use, such as prosthetic control, intelligent remote control, etc. It can be applied to smart home and games. For example, a smart electromyogram bracelet can be worn to control the switch of lights, pause and stop videos; applied to the human-computer interaction in games, such as a real-time tennis game, etc., which can improve the game experience of players. Description of the Drawings

[0056] Figure 1 This is the data preprocessing flowchart of the present invention.

[0057] Figure 2 This is the adaptive update flowchart of the classification model of the present invention. Detailed implementation manners

[0058] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings.

[0059] As Figure 1 、 2 shown, a long-term gesture recognition method based on surface electromyogram signals of the present invention mainly consists of two parts: the first part is data preprocessing, including feature extraction and feature dimensionality reduction; the second part is the adaptive update of the classification model. First is the data preprocessing part. Differential common spatial pattern (DCSP) features are extracted from the gesture data of the existing number of days (usually the data collected on the first day), and then the features are transformed by the non-negative matrix factorization (NMF) algorithm with activation coefficient normalization. Then is the adaptive update part of the classification model. The first repeated experiment of each test day is used as unlabeled samples, and these samples are evaluated and labeled by the clustering and classification self-training (CCST) method, and the qualified samples and labels are used to retrain the classification model together with the existing data.

[0060] The first part: Data preprocessing.

[0061] The steps for extracting DCSP features are as follows:

[0062] Calculate the average covariance matrix of each category signal through formula (1):

[0063]

[0064] where i = 1, 2, 3... c, n is the number of sampling points of the i-th type of action, and X i represents the original electromyogram signal of the i-th category.

[0065] By solving formula (2), find an optimal projection matrix w such that the difference between the i-th category and the j-th category is the largest in the corresponding direction:

[0066] w×R i =λ×w×R j (2)

[0067] A best projection matrix is calculated pairwise for all categories, and all the projection matrices are combined to project the original EMG signal X, as shown in Equation (3):

[0068] A = WX (3)

[0069] Next, the first-order and second-order difference signals of the projected signal are solved and the variances are taken as shown in (4) to obtain A 0 、A 1 、A 2 :

[0070] A 0 = var(A) A 1 = var(ΔA) A 2 = var(Δ 2 A) (4)

[0071] var() represents calculating the variance of the signal, and Δ n represents calculating the nth derivative of the signal, and then through Equation (5) a power exponent transformation is performed to obtain B 0 、B 1 、B 2 , λ is usually taken as 0.1, where Equation (6) combines the features, and finally through (7) logarithmic normalization is performed to obtain the DCSP features:

[0072]

[0073] C = [B 0 B 1 B 2 (6)

[0074] D = log(C / sum(C)) (7)

[0075] In this example, the specific implementation method of converting the DCSP features through the non-negative matrix factorization algorithm NMF with activation coefficient normalization is as follows:

[0076] The method of combining the NMF algorithm with activation coefficient normalization is named the NMFH algorithm, and the DCSP features are converted through the NMFH algorithm to increase the distinguishability of different actions of the DCSP features in long-term gesture recognition;

[0077] After the DCSP features are extracted, the features on the training day are named V, and the features on the test day are named V'. V is decomposed into the product of the collaborative matrix W ∈ R M×K and the activation coefficient matrix H ∈ R K×N by the NMF algorithm. Specifically, the element in the mth row and nth column of V can be decomposed as:

[0078]

[0079] Among them, \(W\) is the muscle synergy matrix, \(H\) is the activation matrix, \(M\) is the number of muscles, \(N\) is the number of feature samples, and \(K\) is the dimensionality of the features after dimensionality reduction. First, \(W\) and \(H\) are initialized as random non-negative matrices. Take \(K = 10\). Then, each element in \(W\) and \(H\) is iteratively solved through equations (9)-(11), where equation (11) is to add the normalization of the activation coefficient \(H\) during the iteration process:

[0080]

[0081]

[0082]

[0083] Test the \(W\) matrix obtained by iteratively fixing the training set in the test set \(V'\). Initialize \(H'\) as a random non-negative matrix, and obtain the test features \(H'\) through equations (10)-(11).

[0084] Part Two: Adaptive Update of the Classification Model

[0085] The specific implementation steps for the adaptive update of the classification model are as follows:

[0086] Step 1: Initialize the classification model and obtain classification labels:

[0087] Train an initial classification model with the samples of the training day. The first repeated experiment of each test day is used as unlabeled samples and marked by the initial classification model to obtain classification labels and classification probabilities. The remaining repeated experiments are used as the test set to verify the performance of the algorithm;

[0088] Step 2: Obtain clustering labels based on fuzzy C-means clustering:

[0089] Obtain the clustering labels of the unlabeled samples through the fuzzy C-means clustering algorithm. Divide the set containing \(n\) data objects into \(c\) classes, and its clustering centers are represented by \(c\) i The objective function is:

[0090]

[0091] In the formula, \(u\) ij takes values between 0 and 1, \(c\) i represents the clustering center of the \(i\)-th class, and \(d\) ij represents the distance between the \(i\)-th clustering center \(c\) i and the \(j\)-th data object \(x\) jThe Euclidean distance between them, where m ∈ [1, +∞) is the weighting exponent used to control the fuzziness of clustering; the larger the value of m, the greater the fuzziness. Under the constraints of the constraint conditions, the Lagrange multiplication is used to obtain the minimum value of Equation (12), and the iterative expressions for the cluster centers and membership degrees are shown in (13) and (14):

[0092]

[0093]

[0094] Take the classification probability of the unlabeled samples in Step 1 as the initial membership degree matrix, and continuously iterate through Equations (13) and (14) to output the cluster centers and membership degree matrix for each round until the objective function (12) reaches the minimum value or exceeds the set maximum number of iterations, at which point the iteration ends; output the membership degree matrix, and use the category corresponding to the maximum value of the membership degree in each unlabeled sample as the clustering label of this unlabeled sample;

[0095] Step 3: Sample evaluation and classifier update:

[0096] Compare the classification labels and clustering labels of each sample, and use the samples with consistent classification labels and clustering labels to update the LDA classifier. The LDA classifier only needs to update two parameters to achieve the update of the classifier; the update steps are as follows:

[0097] μ 2 = rμ 0 +(1 - r)μ 1

[0098] Σ 2 = rΣ 0 +(1 - r)Σ 1

[0099] where u 2 , u 0 and u 1 represent the averages of the updated data, training data, and calibration data respectively; Σ 2 , Σ 0 and Σ 1 represent the covariance matrices of the updated data, training data, and calibration data respectively; r is the balance coefficient, and its value range is between 0 and 1;

[0100] Step 4: Cascade adaptation:

[0101] Adopt a cascaded adaptive method, that is, on the first day of testing, use the training set and the selected data to generate a classification model, and on the second day of testing, use the model updated on the first day and the newly selected data weighted to generate the classification model for the second day. Then, update on the basis of the model updated on the previous day every day thereafter.

[0102] The method of the present invention is applied to the usage process or mode of the product.

[0103] The present invention is a long-term gesture recognition method based on surface electromyogram signals, providing a gesture recognition solution that can be used by users for a long time. The use of this product mainly includes an offline training stage and an online testing stage. In the offline training stage,

[0104] First, the user extracts DCSP features from the existing data (such as the gesture data collected on the first day), then calculates the collaborative matrix W through the NMFH algorithm, and uses the activation coefficient H1 as the training feature. The training feature is used to train the initial LDA classification model. In the online testing stage, DCSP features are also extracted from the signals collected in real time on each testing day. In the NMFH algorithm, W obtained in the training stage is fixed, and H2 is iteratively obtained as the feature. After a small amount of data (one repeated data) is output as an action by the initial classification model, this part of the samples is evaluated and labeled by the CCST algorithm, and updated with the original data to the LDA classification model to obtain a classification model sufficient for the remaining time of the day. Finally, the classification model updated every day is obtained by weighting on the basis of the previous day.

[0105] The above are the preferred embodiments of the present invention. All changes made according to the technical solution of the present invention, when the functions and effects produced do not exceed the scope of the technical solution of the present invention, fall within the protection scope of the present invention.

Claims

1. A long - term gesture recognition method based on surface electromyogram signals, characterized in that, it includes: Data pre - processing: Extract differential common spatial pattern (DCSP) features from gesture data of existing days, and transform the DCSP features through the non - negative matrix factorization algorithm (NMF) with activation coefficient normalization; Adaptive update of the classification model: Take the first repeated experiment of each test day as unlabeled samples, evaluate and label the unlabeled samples using the classification - clustering self - training (CCST) method, and retrain the classification model with the qualified samples, labels and existing data together; The specific implementation of extracting the differential common spatial pattern (DCSP) features is as follows: Calculate the average covariance matrix of each category signal through formula (1): where \(i = 1, 2, 3, \cdots, c\), \(n\) is the number of sampling points of the \(i\)-th category, and \(X\) i represents the original EMG signal of the \(i\)-th category; By solving formula (2), find an optimal projection matrix w such that the difference between the i - th and j - th categories is the largest in the corresponding direction: w×R i = λ×w×R j (2) Calculate an optimal projection matrix for each pair of all categories, combine all projection matrices, and project the original electromyogram signal X, as shown in formula (3): A = wX (3) Next, solve the first-order and second-order difference signals of the projected signal as shown in (4) and take the variance to obtain A 0 , A 1 , A 2 : A 0 = var(A) A 1 = var(ΔA) A 2 = var(Δ 2 A) (4) var() represents calculating the variance of a signal, Δ n represents calculating the nth derivative of a signal, and then performing a power exponential transformation through Equation (5) to obtain B 0 , B 1 , B 2 , where λ is a constant. Among them, Equation (6) combines the features, and finally, logarithmic normalization is performed through (7) to obtain the DCSP feature: C = [B 0 B 1 B 2 (6) D = log(C / sum(C)) (7) The specific implementation of transforming the DCSP features through the non - negative matrix factorization algorithm (NMF) with activation coefficient normalization is as follows: Name the method of combining the NMF algorithm with activation coefficient normalization as the NMFH algorithm, and transform the DCSP features through the NMFH algorithm to increase the distinguishability of different actions of DCSP features in long - term gesture recognition; After DCSP feature extraction, the features on the training day are named V, and the features on the test day are named V'. V is decomposed into the product of a muscle synergy matrix W ∈ R M×K and an activation coefficient matrix H ∈ R K×N by the NMF algorithm. Specifically, the element in the m-th row and n-th column of V is decomposed as follows: Where W is the muscle synergy matrix, H is the activation matrix, M is the number of muscles, N is the number of feature samples, and K is the dimension of the reduced - dimensional features. First, initialize W and H as random non - negative matrices, take K = 10, and then each element in W and H is iteratively solved through formulas (9) - (11), where formula (11) is to add the normalization of the activation coefficient H during the iteration process: Fix the W matrix obtained by the training set iteration of the test set V', initialize H' as a random non - negative matrix, and obtain the test feature H' through formulas (10) - (11).

2. The long - term gesture recognition method based on surface electromyogram signals according to claim 1, characterized in that, the specific implementation steps of the adaptive update of the classification model are as follows: Step 1: Classification model initialization and obtaining classification labels: Train an initial classification model with the samples of the training day. The first repeated experiment of each test day is marked as unlabeled samples by the initial classification model to obtain classification labels and classification probabilities, and the remaining repeated experiments are used as the test set to verify the performance of the algorithm; Step 2: Obtaining clustering labels based on fuzzy C - means clustering: The clustering labels of unlabeled samples are obtained through the fuzzy C-means clustering algorithm. A set containing n data objects is divided into c categories, and its clustering centers are represented by c. The objective function is: i as follows: where u ij ranges from 0 to 1, c i represents the cluster center of the i-th class, d ij represents the Euclidean distance between the i-th cluster center c i and the j-th data object x j , and m ∈ [1, +∞) is the weighted exponent used to control the fuzziness of clustering; The larger the m value, the greater the degree of fuzziness; under the constraints of the constraint conditions, use the Lagrange multiplier method to find the minimum value of formula (12) to obtain the iterative expressions of the clustering center and membership degree as shown in (13) and (14): Take the classification probability of the unlabeled samples in Step 1 as the initial membership matrix, and continuously iterate through Equations (13) and (14) to output the cluster centers and membership matrix for each round until the objective function (12) reaches the minimum value or exceeds the set maximum number of iterations, at which point the iteration ends; output the membership matrix, and use the class corresponding to the maximum membership value in each unlabeled sample as the cluster label for this unlabeled sample. Step 3: Sample evaluation and classifier update: Compare the classification labels and cluster labels of each sample, and use the samples with consistent classification labels and cluster labels to update the LDA classifier. The LDA classifier can be updated by only updating two parameters; the update steps are as follows: μ 2 = rμ 0 +(1 - r)μ 1 Σ 2 = rΣ 0 + (1 - r)Σ 1 where u 2 , u 0 and u 1 represent the average values of updated data, training data, and calibration data, respectively; Σ 2 , Σ 0 and Σ 1 represent the covariance matrices of updated data, training data, and calibration data, respectively; r is a balance coefficient with a value range between 0 and 1; Step 4: Cascade adaptation: Adopt a cascade adaptation method, that is, use the training set and the selected data to generate a classification model on the first day of testing, and use the model updated on the first day and the newly selected data weighted to generate the classification model for the second day of testing. Each subsequent day is updated based on the model updated on the previous day.

Images